IBM Releases Granite 4.2, Its First Open-Weight Reasoning AI Models

IBM has released Granite 4.2, a new family of “reasoning” AI models — models that write out step-by-step reasoning before giving a final answer, rather than jumping straight to a response. The family comes in three sizes, measured by parameters (the internal numbers a model adjusts during training): 3 billion, 8 billion and 30 billion. Earlier Granite models were instruction-following assistants; Granite 4.2 now adds this explicit reasoning ability.

Each model was pre-trained from scratch — trained from the very beginning rather than adapted from an existing model — on roughly 15 trillion tokens (small chunks of text) using a five-phase training process. The final phase extended the model's context window, the amount of text it can consider at once, to 512K tokens.

After pre-training, IBM fine-tuned the models — trained them further on a smaller, targeted dataset — using about 7.2 million examples of chain-of-thought reasoning, instruction-following and tool use, adding up to roughly 100 billion tokens. The models then went through a multi-stage reinforcement learning (RL) pipeline, a training process that rewards better outputs, to sharpen their reasoning further.

The 8B and 30B models also went through an additional stage called agentic RL, teaching them to act as agents: calling outside tools, editing and running code, operating a command-line terminal, and searching the web, all inside sandboxed test environments. Every Granite 4.2 model can switch between a “thinking” mode for harder problems, a faster “non-thinking” mode, and a low-effort thinking mode that spends only a short reasoning budget on easy questions.

All three Granite 4.2 models are released under the Apache 2.0 license as open-weight models — meaning IBM has published the trained model files publicly so anyone can download and use them.